Vera.
Board paper · 4 July 2026 · Confidential · prepared by Andy

Everyone else makes lawyers better at law. Vera runs the firm.

A proposal to redirect Vera from a PCP win-back tool into a catalogue of supervised AI employees for law firm back offices, sold per agent per month, UK first. This paper covers the market evidence, the architecture we'd build on, competitors, pricing, and the build plan.

The ask

Agree the redirection, keep the claims book running as reference customer and funding, and green-light the three-phase build in §8. The verified market window is open now and the platforms are one release cycle behind us.

§1 · Situation

Where we are, and why move.

Vera today is a working, regulatory-grade conversational engine: two-way SMS/email/WhatsApp with dual AI compliance reviewers, a consent ledger, vulnerability screening, signed rule sets and an append-only audit log. It's live on the demo, it has one tenant (BPF, PCP claims), and the people we've shown it to consistently point at the same thing: the compliance safeguards.

The problem isn't the technology, it's the vertical. PCP is getting harder: the FCA redress scheme is contested, CMC economics are tightening, and we're exposed to a single regulator's mood in a single claim type. Meanwhile the engine under Vera, multi-tenant from day one, critic-gated, evidence-first, is far more valuable than the use it's currently put to.

The question this paper answers: what is the biggest thing this engine can become, and is the market really there?

§2 · Thesis

Legal AI is crowded on one side and empty on the other.

Nearly all legal AI money is chasing the practice of law: research, drafting, discovery, case analysis. Harvey, Eve, Spellbook, Luminance. Meanwhile the business of law, onboarding clients, clearing conflicts, chasing documents, collecting bills, keeping clients informed, is served by workflow software that waits for a human to push it.

We verified this with a live, adversarially-checked competitive scan (July 2026, 21 sources, every claim independently verified or killed). The finding is precise: every step of the back-office chain already exists as forms and workflow (Clio Grow, Osprey, Legl), and no vendor anywhere sells autonomous agents that drive it. The gap is not capability, it's autonomy. Workflow software records the work. Nothing does the work.

"You already have workflow software. Vera is the staff member who drives it."

This positioning has a second property that matters to us: back-office operations are not reserved legal activities. No "is the AI practising law?" question, no advice liability. The scariest question in legal AI mostly doesn't apply to the back office, and our supervision architecture (§4) answers the part that does.

§3 · Market

The money has already voted for "AI employees for law firms."

We don't have to argue the category into existence. It's being funded and bought right now, just not in our segment and not in our geography.

$1B+

Eve Legal valuation ($103M Series B, 1,200+ firms). Sells "do law better" to US plaintiff firms. Does not touch the back office. Does not touch the UK.

$15–25k

leap.law charges per month for managed AI agents in Slack for US plaintiff firms, including a daily ops briefing. Proof of demand and a price ceiling.

$3.2M

Caseflood.ai raised in April 2026 for named AI agents that do US personal-injury intake only. The catalogue model works; the scope is narrow.

~9,000

regulated law firms in England & Wales, the overwhelming majority small and mid-sized with no ops staff at all. Partners do the boring bits at partner rates.

The size of the prize, conservatively

Every firm, in every practice area, has onboarding, conflicts, chasing, billing and client updates. The value is genuinely horizontal, which is what makes a catalogue motion coherent. Worked example at our launch pricing (§7): 300 UK firms at an average £1,250/month is £4.5M ARR, roughly 3% of the addressable UK base, before any US or EU expansion and before bundle upsell. leap.law's pricing shows a managed version of the same value selling at 10x our price point in the US today.

Why the window is now

Clio rebranded its assistant to "Manage AI" in early 2026 and Smokeball shipped an agentic "Archie" in May. Both are still human-in-the-loop and scoped to lawyer desk-work, but they are one release cycle away from ops agents. The scan's shelf-life warning is explicit: re-verify by October. We are ahead of them on the two things that are hard to retrofit: autonomy with supervision, and evidence-grade traceability. That lead is real but it is not permanent.

§4 · Product

Four layers. The agents are the least interesting one.

Everyone will have agents within a year. The defensible product is the three layers underneath them, and we already own most of the hardest one.

The Ledger: eight entities, one insight

The whole back office reduces to eight entities. The one nobody else models explicitly is Obligation: "a thing that is owed, and by whom." A document from a client, a signature, a deadline, a task. Practice management systems scatter this across tasks, custom fields and people's heads. Four of our ten digest detections are just queries over this one entity. If the Ledger has a defensible core, this is it.

Client

identity, consent state, vulnerability flags

Matter

status, practice area, key dates, engagement state

Party

anyone related to a matter; feeds conflict checking

CommunicationEvent

every touch, both directions, channel + timestamp

MoneyItem

WIP, invoice, payment, residual balance

Obligation ★

anything owed by anyone: a document, a signature, a deadline

Lead

pre-client enquiries and their follow-up state

AgentAction

the trace record; see the flight recorder below

Supervision: agents work like staff, because firms already trust that model

Law firms already run a supervision model for humans, the SRA requires it: delegated authority from a named supervisor, escalation rules, a complete file. Vera digitises exactly that structure. This isn't marketing language; it is literally the architecture, and most of it is running in the demo today.

  • §i

    Signed, delegated authority

    Each agent operates within a rule set the firm's supervising principal reviews and cryptographically signs (HMAC-sealed, version-tracked). It cannot act outside it. Built; live.

  • §ii

    Two independent AI reviewers

    Every client-facing action passes Counsel (the firm's signed compliance rules) and a separate SRA reviewer (a 7-principle scorecard) before it happens. One redraft on block, then honest escalation to a human. Built; live.

  • §iii

    The autonomy dial, per action type

    Shadow (draft, never send) → draft-for-approval → act-with-audit. Set per action, not globally: "send status updates freely, never chase a bill unsigned." Clio's answer is a binary approve button; ours is consent that is granular and earnable.

  • §iv

    Consent, vulnerability, spend

    Append-only consent ledger (PECR/STOP), two-layer vulnerability screening, per-tenant spend caps and rate limits. Built; live; already validated against FCA-grade outreach.

The flight recorder: every action produces evidence, or it cannot happen

The trace is the atom of the product. One immutable record per action: what the agent knew (the Ledger snapshot), what authority it acted under (rule-set version + the principal's signature), what the reviewers said, what consent permitted it, and what happened. Hash-chained so the log is provably unedited, retained six years like any other file.

Almost every feature is this primitive viewed differently: shadow mode is a trace with execution disabled; receipts are traces aggregated by value; the quarterly audit pack is traces exported as evidence for the SRA, Lexcel/CQS and PII insurers, who now ask firms about AI use. Compliance stops being our cost and becomes an artifact the firm needs anyway.

Strategically: incumbents bolting AI onto twenty-year-old platforms cannot retrofit causation capture. We'd be building it as the substrate everything runs on.

The Loop: the Morning Digest is the product demo, the trojan horse, and the sales pipeline

A firm connects Vera to its PMS in ~20 minutes, read-only. Next morning, the Digest: real findings from their own data, each with a button that activates the agent that fixes it, in shadow mode first. Thirty days later the receipt shows what active agents did and what inactive ones would have done.

Diagnose free, treat paid. It converts our consultancy's Audit → Build → Embed ladder into software, and it's the weapon a future sales team carries: "book a free firm health check" beats any demo deck, because the demo is the firm's own leaked money.

We specified the first ten detections and verified feasibility against Clio's actual API this week: 8 of 10 fully covered, including purpose-built fields (last-activity date, statute-of-limitations flags, overdue-only bill filters, per-matter trust balances). The two gaps are known and designed around: e-signature status has no API (Onboard tracks signatures through its own e-sign flow instead) and Grow's rate limits constrain lead polling at scale.

Independent validation: leap.law's flagship agent is a daily ops briefing with weekly ROI reports, at $15–25k/month, managed. Convergent evolution on the mechanic; we productise it at a tenth of the price.

§5 · Catalogue

Nine agents, tiered by proof and build weight.

Job-description rule stolen from Eve: every agent completes an outcome, never hands off. Every agent reports a monthly receipt.

AgentJob ends atProof / statusTier
IntakeEnquiry triaged, qualified, booked and signed70% of unqualified calls eliminated at Calder & Reid; ~£78k/yr of solicitor time1 · launch
RecoverLapsed lead revived to booked appointmentLive engine (BPF); pull the reactivation number before launch. UK-only until US consent architecture is proven (TCPA)1 · launch
Verify (rename pending)Duplicate claims flagged with per-claim evidence bundle61% duplicate rate found in a motor-finance book; multi-claim engine built and live in demo1 · claims segment
UpdateClient informed before they askComms failure = top source of Legal Ombudsman complaints; a compliance sell, not a convenience sell2 · fast-follow
ChaseDocument or action obtained, matter unblockedCadence engine built (aftercare/reminder system live)2 · fast-follow
BillCash collected or plan agreedReads the ledger; Clio exposes overdue + outstanding-balance endpoints directly2 · fast-follow
ReviewReview posted; close-out loose ends flaggedResidual-balance detection doubles as an Accounts Rules check2 · fast-follow
OnboardClient fully onboarded: ID, AML, engagement letter signed, pushed to PMSThe thesis flagship. Intake state machine exists (Settlement Advice); ID&V via partner (Thirdfort/Amiqus), not built3 · flagship
ConflictClearance report: fuzzy-matched parties + AI summaryOur entity-resolution stack transfers directly; positioned as "Intapp-grade clearance at small-firm prices"3 · flagship

Tiers 1–2 run largely on the chassis we already have and fund the motion. Tier 3 is where the differentiated thesis lives and needs the deep PMS integration. Chase/Update/Bill share one cadence engine differentiated by integration depth, three SKUs from one build.

§6 · Competition

The verified map: everyone is adjacent, nobody is on the square.

From the July 2026 adversarially-verified scan (each claim tested by independent verification passes; four claims killed in the process). This is what survived.

PlayerWhat they actually areDistance from Vera
Clio Manage AI rebranded early 2026Agentic direction, human-in-the-loop desk work: scheduling, bill prep, document analysis. Drafts client comms in assisted mode. "Never finalizes without your consent."Closest platform threat. No onboarding/AML/conflicts/chasing/win-back. One release away; also our #1 distribution channel
Smokeball Archie NG May 2026Agentic multi-step reasoning embedded in Word/Outlook. Desk-work only.Same pattern: agentic tech, lawyer-work scope
Caseflood.ai $3.2M · Apr 2026Named AI agents, US personal-injury intake only. Workflow ends at retainer signing. Site-wide extraction found zero AML/conflicts/letters/chasing scope.Validates the agent-catalogue model; wrong scope, wrong market
leap.law (Legal Automation Partners)Managed AI agents in Slack for US plaintiff firms, incl. daily ops briefing + weekly ROI. $3–5k trial, $15–25k/mo pilots.Closest thesis match alive. Managed service, US, 10x our price. Validates the digest + receipts motion
Legl / Thirdfort / OspreyUK onboarding, ID&V and AML workflow platforms. Thirdfort's May 2026 rebuild adds AI that extracts and interprets source-of-funds data, human-overseen. Legl's only AI is assistive CDD summaries.They are the workflow rails we drive, and partner candidates for ID&V, not head-on competitors
Intapp ConflictsAI conflict triage + AI-generated clearance summaries, shipping since 2024, enterprise-grade.Conflicts is not greenfield. Our claim: Intapp-grade clearance at small-firm prices; segment gap still to verify
Eve Legal $1B+ · USThe category proof. AI for the practice of law: chronologies, demands, discovery, intake for plaintiff firms.Orthogonal by design. Eve does the law; Vera runs the firm. A firm could run both with near-zero overlap. We consciously copied their data-layer moat (Atlas → Ledger) and value-attribution discipline

Scan caveats we're honest about: Amiqus, Verify 365, InfoTrack, MyCase, Actionstep, Filevine, CaseGen and the AI receptionists were fetched but not verified to the same standard; pricing claims produced zero verified survivors (hence §7's caveat); re-scan due October 2026.

§7 · Pricing

Anchor to a caseworker's salary, not a software budget.

The market has organised itself into five price bands. Vera deliberately sits in the gap between "intake platform" and "AI employee contract": richer than the tools below, a fraction of the contracts above, and always compared against a £25–30k FTE, not against software.

BandPrice rangeOccupants
Pure-AI receptionists$49–$400/moSlang.ai, Loman, Rosie, Synthflow-built
AI intake / comms platforms$400–$1,500/moSmith.ai, Gabbi, Case Status, Hona, Intaker
← Vera sits here£750–£1,500 /agent/moWhitespace between the bands
"AI employee" agents$2,000–$5,000/mo11x, Artisan, Caseflood, Eve (effective), leap.law ($15–25k managed)
Per-seat legal copilots$99–$500/seatSpellbook, Alexi, Paxton, Clio Duo add-on
Enterprise legal AI$1,200–$2,400/seatHarvey, Legora ($100k+ contracts)

The working price card, and the motion it implies

Single agent £750/month (launch price), bundles to ~£1,800–£3,000 with expansion discounts, setup £500–£1,500, month-to-month available, US pricing ~1.5–2x UK. The digest is free: it's the pipeline. Published pricing for self-serve tiers; custom tier feeds the Formulaic consultancy (Audit £3.5k → Build → Embed), which also resolves the channel-conflict question: the sales team owns the boundary.

The pricing decides the sales model, not the other way round: a rep at £70k OTE closing £9k-ARR single agents needs 30+ closes a year, unworkable; closing £24–36k-ARR bundles needs 12–15, a normal SMB quota. So: self-serve for single agents and small firms, sales-assisted bundles for 10–25+ fee-earner firms, founder-led for the first five to ten so the team inherits a playbook, not a hypothesis. The health check is the wedge either way.

Caveat, stated plainly: zero pricing claims survived our verification pass. The band logic above is from the earlier pricing-forensics research and bounded by leap.law's verified $15–25k ceiling, but the £750 point is a hypothesis. A dedicated pricing validation is decision #4 in §10.

§8 · Build

We are not starting a company. We are re-pointing one.

The expensive parts, the compliance spine, the conversation engine, the multi-tenant chassis, exist and are live. The build is consolidation and two new layers, not a greenfield.

Already built (live or code-complete)

  • Multi-tenant chassis: tenant-scoped sessions, per-tenant credentials, cost rollups, ~486 tests
  • Dual reviewers: Counsel (signed rule sets, HMAC) + SRA (7-principle scorecard), live on every send
  • Consent ledger, vulnerability screening (2-layer), spend caps, append-only audit
  • Conversation engine: SMS/email/WhatsApp adapters, cadence/reminder system, KB grounding, HITL via Chatwoot
  • Verification engine (the 61% finder): multi-claim, per-claim outcomes, determination PDFs, evidence bundles
  • Intake state machine + magic-link flows (Settlement Advice, to be extracted)
  • Entity resolution (fuzzy matching, dedup) from the location-intelligence work → Conflict
  • Quality-governance pattern (verifier ensembles, regression gates) from et-agent

To build

  • The Ledger: the 8-entity schema + Clio hydration (API feasibility verified this week: 8/10 detections fully covered)
  • Unified supervision service: merge Vera's permission governance and et-agent's quality governance into one chassis layer
  • Hash-chained trace store + the compliance rendering + quarterly audit pack
  • The Digest: ten detections, severity scoring, one-click agent activation
  • Shadow mode as a chassis feature (one build, all agents)
  • Receipts: per-agent monthly value attribution
  • Self-serve onboarding + billing (Stripe, tenant provisioning)
  • Onboard + Conflict flagships (tier 3), ID&V via partner
Phase 1 · productise

Chassis → product

Merge the stacked branches, unify supervision, stand up tenant provisioning and billing. Ship tier-1 agents (Intake, Recover, Verify) on the engine that already runs them. First external tenant.

Phase 2 · the loop

Clio + Ledger + Digest

Clio integration and Ledger hydration, the ten detections, shadow mode, receipts. This is when the health-check motion switches on and the App Directory listing goes in.

Phase 3 · flagships

Onboard + Conflict

The thesis SKUs: full onboarding with partner ID&V and e-sign, and fuzzy-matched conflict clearance. Deepest integration, strongest differentiation, priced accordingly.

Funding logic: the claims book stays on. It is the reference customer ("we run this on our own book"), the proof-point factory, and the cash that buys the runway. Eve cannot say they run their own cases; we can, and it's our best sales asset.

§9 · Risks

What would make this wrong.

§10 · Decisions

What we need from this meeting.

Decision 1 · direction

Approve the redirection

Vera becomes the supervised AI-employee catalogue for law firm back offices, UK first, all practice areas. PCP/BPF continues as tenant one and proof engine, not as the product's identity.

Decision 2 · funding

Keep the claims book on

It funds the build and is the reference customer. Agree the split of attention explicitly so it's a strategy, not a distraction.

Decision 3 · build

Green-light Phase 1

Merge and productise the chassis, unify supervision, first external tenant. The scoping doc exists (VERA-PRODUCT.md); next step is the engineering plan against it.

Decision 4 · evidence

Commission the two studies

Dedicated pricing validation (the one thing our scan couldn't verify) and the October competitive re-scan. Both cheap, both load-bearing.

Decision 5 · sales

Sequence the sales hire

Founder-led for the first 5–10 firms via the health-check motion; hire when the playbook is repeatable. Bundle pricing exists to make that hire's economics work.

Supporting docs

Where the depth lives

VERA-PRODUCT.md (full product spec: architecture, detections, trace schema) and VERA-GTM-MASTER.md (market research, pricing forensics, channel plan). This paper is the synthesis.

Confidential. Prepared 4 July 2026 for the directors of Formulaic. Sources: July 2026 adversarially-verified competitive scan (21 sources), Clio API feasibility spike, pricing-forensics research, live Vera demo.
Vera. · Formulaic